From 843ccdb806055b9139240b95cf49d8654554f21a Mon Sep 17 00:00:00 2001
From: Mina Choi <mnchoi@o2o.kr>
Date: Wed, 27 May 2026 13:27:39 +0900
Subject: [PATCH] =?UTF-8?q?=EB=B8=8C=EB=9E=9C=EB=93=9C=20=EC=9E=90?=
 =?UTF-8?q?=EC=82=B0(=EB=A1=9C=EA=B3=A0/=EC=83=89=EC=83=81)=C2=B7=EC=B1=84?=
 =?UTF-8?q?=EB=84=90=20=EB=A1=9C=EA=B3=A0=20Vision=20=EB=B6=84=EC=84=9D=20?=
 =?UTF-8?q?=EC=B6=94=EA=B0=80?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- color_extractor: 홈페이지 HTML/CSS에서 로고 URL·브랜드 hex 추출
- vision: Gemini Vision 로고 묘사·채널 로고 일치 평가
- youtube: 채널 profileImage 추출 / firecrawl: clinic_info 추출 보정

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 app/integrations/color_extractor.py | 250 ++++++++++++++++++++++++++++
 app/integrations/firecrawl.py       |   3 +-
 app/integrations/vision.py          | 171 +++++++++++++++++++
 app/integrations/youtube.py         |   2 +
 4 files changed, 425 insertions(+), 1 deletion(-)
 create mode 100644 app/integrations/color_extractor.py
 create mode 100644 app/integrations/vision.py

diff --git a/app/integrations/color_extractor.py b/app/integrations/color_extractor.py
new file mode 100644
index 0000000..806d8cd
--- /dev/null
+++ b/app/integrations/color_extractor.py
@@ -0,0 +1,250 @@
+"""홈페이지 HTML/CSS에서 hex 색상 직접 추출 + 빈도 기반 brand palette 산출.
+
+Vision LLM에 의존하지 않고 페이지의 실제 CSS 값을 정규식으로 잡음.
+로고만 분석하는 Vision보다 사이트 전체 컬러 시스템 (primary/secondary/background/text)을 더 정확히 추출.
+"""
+import logging
+import re
+import ssl
+from collections import Counter
+from urllib.parse import urljoin, urlparse
+import httpx
+
+logger = logging.getLogger(__name__)
+
+
+def _make_ssl_context() -> ssl.SSLContext:
+    """오래된 한국 의료 사이트들이 SSL DH_KEY_TOO_SMALL / cipher 약함 등으로 차단되는 문제 우회.
+    보안 등급 1로 낮춤 + cert 검증 유지."""
+    ctx = ssl.create_default_context()
+    try:
+        ctx.set_ciphers("DEFAULT@SECLEVEL=1")
+    except ssl.SSLError:
+        pass
+    return ctx
+
+
+async def _fetch_html(url: str, timeout: float = 20.0) -> tuple[int, str]:
+    """SSL/검증 단계별 fallback으로 HTML 받기. 그랜드/톡스앤필 같은 oldsite 대응."""
+    headers = {"User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36"}
+    # 1차: 표준 검증
+    try:
+        async with httpx.AsyncClient(timeout=timeout, follow_redirects=True, headers=headers) as c:
+            r = await c.get(url)
+            return r.status_code, r.text
+    except (httpx.ConnectError, httpx.ReadError, ssl.SSLError) as e:
+        logger.info("[fetch] %s standard SSL failed: %s — fallback to weak cipher", url, e)
+    # 2차: 약한 cipher 허용
+    try:
+        async with httpx.AsyncClient(timeout=timeout, follow_redirects=True, headers=headers, verify=_make_ssl_context()) as c:
+            r = await c.get(url)
+            return r.status_code, r.text
+    except (httpx.ConnectError, httpx.ReadError, ssl.SSLError) as e:
+        logger.info("[fetch] %s weak cipher failed: %s — fallback to verify=False", url, e)
+    # 3차: SSL 검증 끔 (host mismatch 등)
+    try:
+        async with httpx.AsyncClient(timeout=timeout, follow_redirects=True, headers=headers, verify=False) as c:
+            r = await c.get(url)
+            return r.status_code, r.text
+    except Exception as e:
+        logger.warning("[fetch] %s all fallbacks failed: %s", url, e)
+        return 0, ""
+
+LOGO_IMG_PATTERNS = [
+    # 1) <img class="...logo..." src="...">
+    re.compile(r'<img[^>]*\bclass=["\'][^"\']*\blogo\b[^"\']*["\'][^>]*\bsrc=["\']([^"\']+)["\']', re.IGNORECASE),
+    # 2) <img src="..." class="...logo...">
+    re.compile(r'<img[^>]*\bsrc=["\']([^"\']+)["\'][^>]*\bclass=["\'][^"\']*\blogo\b[^"\']*["\']', re.IGNORECASE),
+    # 3) <img id="...logo..." src="...">
+    re.compile(r'<img[^>]*\bid=["\'][^"\']*\blogo\b[^"\']*["\'][^>]*\bsrc=["\']([^"\']+)["\']', re.IGNORECASE),
+    # 4) <img alt="...logo..." src="...">
+    re.compile(r'<img[^>]*\balt=["\'][^"\']*\blogo\b[^"\']*["\'][^>]*\bsrc=["\']([^"\']+)["\']', re.IGNORECASE),
+    # 5) <a/h1 class="logo"><...nested...><img src="...">
+    re.compile(r'<(?:a|h[1-6]|div|span)[^>]*\b(?:class|id)=["\'][^"\']*\blogo\b[^"\']*["\'][^>]*>(?:[^<]|<(?!img))*<img[^>]*\bsrc=["\']([^"\']+)["\']', re.IGNORECASE | re.DOTALL),
+    # 6) inline background-image: <a/div class="logo" style="background-image: url(...)">
+    re.compile(r'<(?:a|div|span|h[1-6])[^>]*\b(?:class|id)=["\'][^"\']*\blogo\b[^"\']*["\'][^>]*\bstyle=["\'][^"\']*background(?:-image)?\s*:\s*url\(\s*["\']?([^"\')\s]+)', re.IGNORECASE),
+    # 7) inline background-image: <a/div style="background-image: url(...)" class="logo">  (속성 순서 반대)
+    re.compile(r'<(?:a|div|span|h[1-6])[^>]*\bstyle=["\'][^"\']*background(?:-image)?\s*:\s*url\(\s*["\']?([^"\')\s]+)[^"\']*["\'][^>]*\b(?:class|id)=["\'][^"\']*\blogo\b', re.IGNORECASE),
+    # 8) src 자체에 "logo" 포함 (header_logo.png, brand-logo.svg 등)
+    re.compile(r'<img[^>]*\bsrc=["\']([^"\']*\blogo\b[^"\']*\.(?:png|svg|jpe?g|webp)[^"\']*)["\']', re.IGNORECASE),
+    # 9) <header>...<img src="..."> (헤더 영역 첫 img)
+    re.compile(r'<header\b[^>]*>(?:[^<]|<(?!img))*<img[^>]*\bsrc=["\']([^"\']+\.(?:png|svg|jpe?g|webp)[^"\']*)["\']', re.IGNORECASE | re.DOTALL),
+    # 10) <nav>...<img src="..."> (nav 영역 첫 img)
+    re.compile(r'<nav\b[^>]*>(?:[^<]|<(?!img))*<img[^>]*\bsrc=["\']([^"\']+\.(?:png|svg|jpe?g|webp)[^"\']*)["\']', re.IGNORECASE | re.DOTALL),
+    # 11) Open Graph image (대표 이미지) - 최후 fallback
+    re.compile(r'<meta[^>]*\bproperty=["\']og:image["\'][^>]*\bcontent=["\']([^"\']+)["\']', re.IGNORECASE),
+    re.compile(r'<meta[^>]*\bcontent=["\']([^"\']+)["\'][^>]*\bproperty=["\']og:image["\']', re.IGNORECASE),
+]
+
+# CSS 파일에서 .logo { background-image: url(...) } 추출용
+LOGO_CSS_PATTERN = re.compile(
+    r'\.[\w-]*\blogo\b[\w-]*\s*(?:,\s*\.[\w-]+\s*)*\{[^}]*background(?:-image)?\s*:\s*url\(\s*["\']?([^"\')\s]+)',
+    re.IGNORECASE | re.DOTALL,
+)
+
+
+def find_logo_url_in_html(html: str, base_url: str, css_texts: list[str] | None = None) -> str | None:
+    """HTML에서 logo URL 찾기. class/id/alt → 부모 + 중첩 img → background-image → src에 logo → header/nav → og:image 순."""
+    for pat in LOGO_IMG_PATTERNS:
+        for m in pat.finditer(html):
+            src = m.group(1)
+            if not src or src.startswith("data:"):
+                continue
+            if re.search(r"(blank|spacer|pixel|transparent|1x1)\b", src, re.IGNORECASE):
+                continue
+            return urljoin(base_url, src)
+    # 외부 CSS에서 .logo background-image 추출
+    for css in (css_texts or []):
+        m = LOGO_CSS_PATTERN.search(css)
+        if m:
+            src = m.group(1)
+            if src and not src.startswith("data:"):
+                return urljoin(base_url, src)
+    return None
+
+
+HEX6 = re.compile(r"#([0-9a-fA-F]{6})\b")
+HEX3 = re.compile(r"#([0-9a-fA-F]{3})\b(?![0-9a-fA-F])")
+RGB  = re.compile(r"rgba?\(\s*(\d{1,3})\s*,\s*(\d{1,3})\s*,\s*(\d{1,3})\s*(?:,\s*[\d.]+\s*)?\)")
+CSS_VAR_HEX = re.compile(r"--[\w-]+\s*:\s*(#[0-9a-fA-F]{3,8})", re.IGNORECASE)
+CSS_LINK = re.compile(r'<link[^>]+rel=["\']stylesheet["\'][^>]+href=["\']([^"\']+)["\']', re.IGNORECASE)
+STYLE_BLOCK = re.compile(r"<style[^>]*>(.*?)</style>", re.IGNORECASE | re.DOTALL)
+
+# 무채색·아주 흔한 노이즈 컬러 (이런 건 brand color로 잡지 않음)
+NOISE = {
+    "#ffffff", "#000000", "#fff", "#000",
+    "#333", "#222", "#111", "#444", "#555", "#666", "#777", "#888", "#999",
+    "#aaa", "#bbb", "#ccc", "#ddd", "#eee", "#f0f0f0", "#f5f5f5", "#fafafa",
+}
+
+
+def _normalize(hex_str: str) -> str:
+    h = hex_str.lstrip("#").lower()
+    if len(h) == 3:
+        h = "".join(c * 2 for c in h)
+    if len(h) == 8:
+        h = h[:6]
+    return f"#{h}"
+
+
+def _rgb_to_hex(r: int, g: int, b: int) -> str:
+    return f"#{r:02x}{g:02x}{b:02x}"
+
+
+def _hex_to_rgb(h: str) -> tuple[int, int, int]:
+    h = h.lstrip("#")
+    return int(h[0:2], 16), int(h[2:4], 16), int(h[4:6], 16)
+
+
+def _distance(a: str, b: str) -> float:
+    ar, ag, ab = _hex_to_rgb(a)
+    br, bg, bb = _hex_to_rgb(b)
+    return ((ar - br) ** 2 + (ag - bg) ** 2 + (ab - bb) ** 2) ** 0.5
+
+
+def _is_grayscale(h: str, tol: int = 12) -> bool:
+    r, g, b = _hex_to_rgb(h)
+    return max(r, g, b) - min(r, g, b) < tol
+
+
+def _extract_hex(text: str) -> list[str]:
+    """텍스트에서 모든 hex 색상 추출 (정규화)."""
+    out: list[str] = []
+    out.extend(_normalize(m.group(0)) for m in HEX6.finditer(text))
+    out.extend(_normalize(m.group(0)) for m in HEX3.finditer(text))
+    for m in RGB.finditer(text):
+        r, g, b = int(m.group(1)), int(m.group(2)), int(m.group(3))
+        if 0 <= r <= 255 and 0 <= g <= 255 and 0 <= b <= 255:
+            out.append(_rgb_to_hex(r, g, b))
+    return out
+
+
+def _cluster(colors: Counter, threshold: float = 25.0) -> list[tuple[str, int]]:
+    """비슷한 색은 묶음. 가장 빈도 높은 색을 대표로."""
+    ranked = colors.most_common()
+    clusters: list[tuple[str, int]] = []
+    for color, count in ranked:
+        merged = False
+        for i, (rep, rep_count) in enumerate(clusters):
+            if _distance(color, rep) < threshold:
+                clusters[i] = (rep, rep_count + count)
+                merged = True
+                break
+        if not merged:
+            clusters.append((color, count))
+    return clusters
+
+
+async def _fetch_html_and_css(homepage_url: str, max_css_files: int = 8) -> tuple[str, list[str]]:
+    """홈페이지 HTML + 외부 CSS(Top N)를 한 번에 fetch. 로고/색상 추출이 사이트를 중복으로 긁지 않도록 공유.
+    _fetch_html이 SSL 약함/host mismatch까지 fallback 처리. 실패 시 ("", [])."""
+    status, html = await _fetch_html(homepage_url)
+    if status != 200 or not html:
+        logger.warning("[color_extractor] homepage fetch failed status=%s url=%s", status, homepage_url)
+        return "", []
+    css_texts: list[str] = []
+    for css_href in CSS_LINK.findall(html)[:max_css_files]:
+        cstatus, ctext = await _fetch_html(urljoin(homepage_url, css_href), timeout=15.0)
+        if cstatus == 200 and ctext:
+            css_texts.append(ctext)
+    return html, css_texts
+
+
+def _colors_from_text(html: str, css_texts: list[str], source_url: str = "") -> dict:
+    """이미 받아온 HTML + CSS 텍스트에서 hex 빈도 분석 → primary/accent/text + palette. (fetch 없음, 순수 계산)"""
+    # 1. HTML 내 <style> 블록 + 통째(inline style="color:#...") + 외부 CSS
+    all_text_chunks: list[str] = list(STYLE_BLOCK.findall(html))
+    all_text_chunks.append(html)
+    all_text_chunks.extend(css_texts)
+
+    # 2. 모든 hex 추출 (NOISE 제외)
+    counter: Counter = Counter()
+    for text in all_text_chunks:
+        for color in _extract_hex(text):
+            if color in NOISE:
+                continue
+            counter[color] += 1
+
+    if not counter:
+        logger.info("[color_extractor] no colors extracted from %s", source_url)
+        return {}
+
+    # 3. 비슷한 색 클러스터링
+    clustered = _cluster(counter)
+
+    # 4. primary = 빈도 높은 채도 있는 색 / accent = 두번째 채도 있는 색 / text = 빈도 높은 무채색
+    chromatic = [c for c, _ in clustered if not _is_grayscale(c)]
+    grayscale = [c for c, _ in clustered if _is_grayscale(c)]
+
+    palette_top = clustered[:8]
+    palette = [{"name": f"색상 {i+1}", "hex": h, "usage": f"빈도 {n}"} for i, (h, n) in enumerate(palette_top)]
+
+    return {
+        "brand_colors": {
+            "primary": chromatic[0] if chromatic else None,
+            "accent": chromatic[1] if len(chromatic) > 1 else None,
+            "text": grayscale[0] if grayscale else None,
+        },
+        "color_palette": palette,
+        "extracted_from": "html+css",
+    }
+
+
+async def extract_brand_colors_from_site(homepage_url: str, max_css_files: int = 8) -> dict:
+    """홈페이지 HTML + 외부 CSS fetch → hex 색상 빈도 분석 → primary/accent/text + palette 5종."""
+    html, css_texts = await _fetch_html_and_css(homepage_url, max_css_files)
+    if not html:
+        return {}
+    return _colors_from_text(html, css_texts, homepage_url)
+
+
+async def extract_brand_assets_from_site(homepage_url: str, max_css_files: int = 8) -> dict:
+    """사이트를 한 번만 fetch해서 로고 URL과 brand 색상을 함께 추출.
+    반환: {"logo_url": str | None, "colors": {brand_colors, color_palette, ...} | {}}"""
+    html, css_texts = await _fetch_html_and_css(homepage_url, max_css_files)
+    if not html:
+        return {"logo_url": None, "colors": {}}
+    return {
+        "logo_url": find_logo_url_in_html(html, homepage_url, css_texts=css_texts),
+        "colors": _colors_from_text(html, css_texts, homepage_url),
+    }
diff --git a/app/integrations/firecrawl.py b/app/integrations/firecrawl.py
index 7daa866..994daba 100644
--- a/app/integrations/firecrawl.py
+++ b/app/integrations/firecrawl.py
@@ -76,7 +76,7 @@ class FirecrawlClient:
                 "url": url,
                 "formats": ["json", "links"],
                 "jsonOptions": {
-                    "prompt": "Extract: clinic name (Korean), clinic name (English), address, phone with dash format, business hours, slogan, services offered, doctors with name/title/specialty, brand identity (primary/accent/background/text colors in hex, heading/body fonts, logo URL, favicon URL)",
+                    "prompt": "Extract: clinic name (Korean), clinic name (English), address, phone with dash format, business hours, slogan, services offered, doctors with name/title/specialty, brand identity (primary/accent/background/text colors in hex, heading/body fonts, logo URL from the actual header/main <img> src, og:image from <meta property='og:image'> content, favicon URL)",
                     "schema": {
                         "type": "object",
                         "properties": {
@@ -119,6 +119,7 @@ class FirecrawlClient:
                                     "headingFont":     {"type": "string"},
                                     "bodyFont":        {"type": "string"},
                                     "logoUrl":         {"type": "string"},
+                                    "ogImage":         {"type": "string"},
                                     "faviconUrl":      {"type": "string"},
                                 },
                             },
diff --git a/app/integrations/vision.py b/app/integrations/vision.py
new file mode 100644
index 0000000..67371a2
--- /dev/null
+++ b/app/integrations/vision.py
@@ -0,0 +1,171 @@
+"""Gemini Vision — 로고/브랜드 비주얼 자동 분석 (OpenAI 호환 모드).
+
+정확한 hex 색상은 color_extractor가 CSS에서 직접 뽑음 (Vision은 근사값밖에 못 냄).
+Vision은 사람이 봐야 알 수 있는 정성 정보 — 심볼 형태/워드마크/톤 — 를 담당.
+"""
+import base64
+import json
+import logging
+import re
+import httpx
+from openai import AsyncOpenAI
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_MODEL = "gemini-2.5-flash"
+
+
+class VisionClient:
+    """Gemini Vision을 OpenAI 호환 endpoint로 호출. GEMINI_API_KEY만 필요."""
+
+    def __init__(self, api_key: str, model: str = DEFAULT_MODEL, timeout: float = 30.0, max_retries: int = 2):
+        self.client = AsyncOpenAI(
+            api_key=api_key,
+            base_url="https://generativelanguage.googleapis.com/v1beta/openai/",
+            timeout=timeout,
+            max_retries=max_retries,
+        )
+        self.model = model
+
+    @staticmethod
+    def _extract_json(text: str) -> dict | None:
+        if not text:
+            return None
+        m = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", text, re.DOTALL)
+        if m:
+            try:
+                return json.loads(m.group(1))
+            except json.JSONDecodeError:
+                pass
+        m = re.search(r"\{.*\}", text, re.DOTALL)
+        if m:
+            try:
+                return json.loads(m.group(0))
+            except json.JSONDecodeError:
+                return None
+        return None
+
+    @staticmethod
+    async def _fetch_as_data_url(url: str) -> str | None:
+        """Gemini는 URL 직접 fetch가 막힌 호스트가 많아 base64 인라인으로 변환.
+        + 'image does not exist' 같은 placeholder 이미지 거부 (작은 bytes / 잘못된 content-type)."""
+        try:
+            async with httpx.AsyncClient(timeout=15.0, follow_redirects=True) as c:
+                resp = await c.get(url)
+                if resp.status_code != 200:
+                    logger.warning("[vision] fetch %s status=%s", url, resp.status_code)
+                    return None
+                mime = resp.headers.get("content-type", "").split(";")[0].strip()
+                # 실제 이미지가 아니면 거부 (HTML 페이지가 404 대신 200으로 리다이렉트 되는 경우)
+                if not mime.startswith("image/"):
+                    logger.warning("[vision] %s not an image (content-type=%s)", url, mime)
+                    return None
+                size = len(resp.content)
+                if size < 500:
+                    logger.warning("[vision] %s too small (%d bytes) — likely placeholder", url, size)
+                    return None
+                b64 = base64.b64encode(resp.content).decode("ascii")
+                return f"data:{mime};base64,{b64}"
+        except Exception as e:
+            logger.warning("[vision] fetch error %s: %s", url, e)
+            return None
+
+    async def _ask(self, image_urls: list[str], prompt: str, max_tokens: int = 4000) -> dict | None:
+        content: list[dict] = []
+        for u in image_urls:
+            if not u:
+                continue
+            data_url = await self._fetch_as_data_url(u)
+            if not data_url:
+                continue
+            content.append({"type": "image_url", "image_url": {"url": data_url}})
+        if not any(c.get("type") == "image_url" for c in content):
+            logger.warning("[vision] no images could be fetched")
+            return None
+        content.append({"type": "text", "text": prompt})
+
+        try:
+            resp = await self.client.chat.completions.create(
+                model=self.model,
+                messages=[{"role": "user", "content": content}],
+                max_tokens=max_tokens,
+            )
+            choice = resp.choices[0]
+            if choice.finish_reason != "stop":
+                logger.warning("[vision] unexpected finish_reason=%s", choice.finish_reason)
+            return self._extract_json(choice.message.content or "")
+        except Exception as e:
+            logger.warning("[vision] error: %s", e)
+            return None
+
+    async def analyze_brand_assets(
+        self,
+        logo_url: str | None,
+        homepage_url: str | None,
+        additional_images: list[str] | None = None,
+    ) -> dict:
+        """로고 이미지를 보고 정성 분석. 정확한 hex는 color_extractor가 따로 처리하므로 여기선 안 뽑음."""
+        urls = [u for u in [logo_url] + list(additional_images or []) if u]
+        if not urls:
+            return {}
+        prompt = (
+            "당신은 브랜드 로고 시각 분석가입니다. 첨부된 이미지(첫 번째가 병원의 대표 로고)를 보고 "
+            "아래 JSON 스키마로만 응답하세요. 코드펜스 없이 순수 JSON만 출력.\n"
+            "{\n"
+            '  "logo_description": "로고를 1~2문장으로 설명 (심볼 형태 + 워드마크 + 전반적 톤). 예: \'둥근 잎사귀를 감싼 추상 심볼에 세리프 한글 워드마크, 차분하고 고급스러운 톤\'",\n'
+            '  "logo_style": "minimal | illustrative | typographic | abstract 중 하나",\n'
+            '  "has_symbol": "심볼/아이콘이 있으면 true, 글자만 있으면 false (boolean)",\n'
+            '  "logo_symbol": "심볼이 묘사하는 대상 (예: \'잎사귀\', \'추상 곡선\'). 없으면 빈 문자열",\n'
+            '  "logo_text": "로고에 보이는 워드마크 텍스트 그대로 (한글/영문). 없으면 빈 문자열",\n'
+            '  "logo_colors_desc": "로고에 쓰인 색감을 사람이 부르는 이름으로 서술 (예: \'딥네이비 + 골드\'). 정확한 hex는 출력하지 말 것"\n'
+            "}\n"
+            "주의: 색상 hex 값이나 logo URL 같은 필드는 출력하지 마세요 (별도 추출 로직이 처리)."
+        )
+        result = await self._ask(urls, prompt)
+        if not result:
+            return {}
+        # logo_images는 우리가 직접 채움 (Vision은 묘사만)
+        result["logo_images"] = {"circle": None, "horizontal": logo_url, "korean": None}
+        return result
+
+    async def describe_channel_logos(
+        self,
+        official_logo_url: str | None,
+        channel_logos: list[dict],
+    ) -> dict | None:
+        """채널별 프로필 이미지(로고)를 보고 각각 설명 + 공식 로고와 일치 여부 평가.
+        channel_logos: [{"channel": "Instagram", "url": "..."}, ...]
+        반환: {"channel_logos": [{"channel","logo_description","is_official"}], "inconsistency_summary", "recommendation"}"""
+        items = [c for c in channel_logos if c.get("url")]
+        if not items:
+            return None
+
+        # 공식 로고가 있으면 맨 앞에 두고 기준으로 삼음
+        urls: list[str] = []
+        if official_logo_url:
+            urls.append(official_logo_url)
+        urls.extend(c["url"] for c in items)
+        channel_order = ", ".join(c.get("channel", "?") for c in items)
+
+        if official_logo_url:
+            header = (
+                "첨부 이미지 중 **첫 번째가 이 병원의 공식 로고**입니다. "
+                f"이어지는 이미지들은 채널별 프로필 이미지이며 순서는: {channel_order}.\n"
+                "각 채널 로고를 1문장으로 설명하고, 공식 로고(첫 번째)와 일치하면 is_official=true, "
+                "비공식 변형/모델사진/다른 이미지면 false로 평가하세요.\n"
+            )
+        else:
+            header = (
+                f"첨부 이미지는 한 병원의 채널별 프로필 이미지입니다. 순서: {channel_order}.\n"
+                "각 채널 로고를 1문장으로 설명하세요 (공식 로고 기준이 없으므로 is_official은 판단 가능하면만).\n"
+            )
+        prompt = (
+            header
+            + "아래 JSON으로만 응답 (코드펜스 없이 순수 JSON):\n"
+            "{\n"
+            '  "channel_logos": [{"channel": "...", "logo_description": "...", "is_official": true}],\n'
+            '  "inconsistency_summary": "채널 간 로고 일관성 1~2문장 요약",\n'
+            '  "recommendation": "통합 권고 1문장"\n'
+            "}"
+        )
+        return await self._ask(urls, prompt)
diff --git a/app/integrations/youtube.py b/app/integrations/youtube.py
index 734f142..0a5c1a9 100644
--- a/app/integrations/youtube.py
+++ b/app/integrations/youtube.py
@@ -88,9 +88,11 @@ class YouTubeClient:
         ch = raw["channel"]
         stats = ch.get("statistics", {})
         snippet = ch.get("snippet", {})
+        thumbs = snippet.get("thumbnails", {})
         return {
             "channelId": raw["channelId"],
             "channelName": snippet.get("title"),
+            "profileImage": (thumbs.get("high") or thumbs.get("medium") or thumbs.get("default") or {}).get("url"),
             "handle": snippet.get("customUrl"),
             "description": snippet.get("description", ""),
             "publishedAt": snippet.get("publishedAt"),